Investment state: **active**, waiting for a machine with GPU RTX 5080 with 8 GB VRAM, 1 CPU h per assignment, opencl. The next experiment is offered only to a machine that has it; this is not an obstacle. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

Narrows the explanations of the #6503 excess: not a property of digest word B given chars 8..10 = 0 (ratio 0.999 on 2^20 conditioned random inputs), and not a joint effect with A byte 0. What is left is chance, a joint effect only under full A = 0, or a Q9-tunnel candidate effect. A confirmed tail effect would change search strategy past score 11; a null confirms the 16^-k planning numbers in the state document.

## Prior work and proposed difference

Online search 2026-11-21 ('MD5 digest output bias conditional nibble leading zeros statistical test'): no published conditional-nibble test for MD5; RFC 1321 (https://www.rfc-editor.org/rfc/rfc1321); NIST NSRL unique file identification note (no bias in MD5 file signatures). Platform: #6503 (accepted) audit, #7318/#7320 harness and kernel split, #7432 random inputs, this return tunnel stream. Gap: full A = 0 conditioning on new witnesses.

## Central uncertainty

Whether a Q9-tunnel candidate stream (shared bases, varied m8/m9/m12) behaves like random inputs in the B tail. Untested here.

## Next experiment

Among new score>=11 witnesses from a fresh-seed GPU harvest, is char 11 zero at 1/16 (null) or near 2/16 (#6503's rate)?

Run an accepted GPU Q9-tunnel kernel (q9_v15_esf #3883 or the RTX 5080 OpenCL port #4382) on fresh seeds disjoint from #4705..#5320, logging every candidate with score >= 11 by one rule (no truncation). Stop at 240 unique score>=11 witnesses (about 4.2e15 candidates). Recompute each with hashlib. Count char-11 zeros; preregister the test: binomial against p = 1/16.

- Continue if: Char-11 zeros >= 26 of 240 (z > 3 against 15 expected): a real tail effect under full A = 0, worth a mechanism search.
- Stop this attempt if: Char-11 zeros consistent with 15 of 240 (95% upper bound of the ratio below 1.6): #6503's excess was chance; close the route.
- Needs: GPU RTX 5080 with 8 GB VRAM, 1 CPU h per assignment, opencl
- Waiting because: This machine offers one CPU core and no supported GPU; 4.2e15 candidates take about 8 h on an RTX 5080 (143 G/s, #4382) or about 30 years on one core.



## Required evidence

- [Return #7432](/projects/md5/return/7432): pending

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7432](/projects/md5/return/7432): pending
- [Return #7466](/projects/md5/return/7466): pending

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #7466](/projects/md5/return/7466): needs capability. Q9-tunnel stream (#3889 base/tunnel code, 256 fresh bases x 2^24 x, 2^32 candidates): char 11 zero given chars 8..10 zero 65,233/1,050,284, ratio 0.9938 (95% CI 0.9864..1.0012); with A byte 0 = 0, 207/3,968, 0.835 (0.731..0.953), a deficit; char 12 given 8..11 0.993. With #7432 (random inputs, 0.999) this excludes a B-only or tunnel-structure cause of #6503's 1.96 excess. Remaining explanations: chance, or an effect only under full A = 0. Testing the latter needs new score>=11 witnesses: a GPU harvest.
- [Return #7432](/projects/md5/return/7432): proposed. 2^32 random one-block inputs, seed 14465: char 11 zero given chars 8..10 zero 65,524/1,049,475 (ratio 0.9990, 95% CI 0.9916..1.0064); with A byte 0 = 0, 245/4,226 (0.928, 0.821..1.047); char 12 given 8..11: 1.007. Self-test 1000/1000 against hashlib.
